Monte Carlo vs Historical Sequence Testing: Which Is More Honest?

What each method does

Monte Carlo simulation generates thousands of synthetic return sequences from an assumed probability distribution. The most common setup: assume returns are normally distributed with a mean of 6% real and a standard deviation of 15%. Generate 10,000 different 50-year paths. Run your FIRE plan against each. Report the percentage that succeed.

Historical sequence testing uses the actual order returns occurred in the past. For each possible starting year in the data, run your plan against the exact subsequent sequence. There are roughly 100 overlapping 50-year windows in the Shiller dataset; you get 100 simulations.

Both methods estimate the same thing — how robust is your plan? — but they make different assumptions about how markets behave.

Where they disagree

The biggest disagreement is on sequence risk. Monte Carlo assumes returns are independent (today's return tells you nothing about tomorrow's). Historical data shows they aren't:

• Short-term momentum: returns are positively autocorrelated over 3–12 months. Up markets tend to keep going up.
• Long-term mean reversion: returns are negatively autocorrelated over 3–7 years. Strong decades are usually followed by weaker ones.
• Volatility clustering: high-volatility periods cluster together. 2008 didn't come out of nowhere — it followed a period of building stress.

Pure Monte Carlo with normal returns misses all three. As a result, it tends to be too optimistic about long-horizon plans because it never generates the kind of multi-year grinding underperformance that 1966 and 2000 actually produced.

What Monte Carlo does well

Don't dismiss it. Monte Carlo has real strengths:

• Large sample size. You can run 100,000 paths, which gives statistically clean tail estimates.
• Customisable assumptions. You can explicitly model an expected-return shock — say, what happens if future equity returns are 3% real instead of 6%?
• Independence from the historical record. Useful if you think the future will differ structurally from the past (e.g. new monetary regime).

For asking "what if returns are systematically lower than history suggests", Monte Carlo is the right tool.

What historical testing does well

• Captures real patterns. Momentum, mean reversion, regime shifts — all present in the data, all automatically reflected.
• Concrete worst cases. You can name the bad cohorts: 1929, 1966, 2000. That's easier to plan against than an abstract "1% tail" from a distribution.
• Forces realism. If a plan fails against 1966, you can't argue with that — it would have actually failed.

For asking "would my plan have survived the historical record", historical testing is the right tool.

How FIRE Wealth OS uses both

We default to historical sequence testing because we think it's the more honest answer for the "what's the safe withdrawal rate" question. The output shows you which actual historical cohorts your plan would have broken in, which is more actionable than a percentage.

But for stress testing — "what if future returns are 30% worse than history?" — Monte Carlo gives a cleaner answer. Our premium tier includes both, so you can compare the two estimates for your specific plan.

The pragmatic conclusion: trust historical testing for the headline number, use Monte Carlo to test pessimistic assumptions about the future. If both methods agree your plan is robust, you have a robust plan.

The biggest mistake either approach can make

Treating either output as a precise answer. Neither method tells you "your plan has a 96.3% chance of success". They give you a distribution under specific assumptions, and the assumptions are doing most of the work.

Sensible use of either method:

• Take the worst 10% of outcomes seriously
• Don't optimise to the 95th percentile, because that's where false precision lives
• Re-run with multiple assumptions and see how the answer moves

Open the simulator and check the spread of outcomes. The spread is the actual insight; the median is just one slice of it.